JP5775425B2 - Purchase data analysis apparatus, method, and program - Google Patents

Description

  The present invention relates to a purchase data analysis apparatus, method, and program, and more particularly, to a purchase data analysis apparatus, method, and program for estimating a model for selecting a product purchased by a user based on purchase data.

  When a user purchases a product, information about all the products is not collected and evaluated. The user selects the final purchase product by narrowing down the products (consideration set) to be considered from all the products (available set), and evaluating and examining the products in the consideration set.

  In addition, a purchased product selection model that models purchase behavior in which a user selects a product to be purchased has been proposed (Non-Patent Document 1).

Tomoharu Iwata, Junji Watanabe, Takeshi Yamada, Nobuo Ueda, "Topic Tracking Model for Purchasing Behavior Analysis", IEICE, Vol.J93-D, No.6, pp.978-987, 2010 (issue date) : June 1, 2010).

  In the said nonpatent literature 1, there exists a problem that the purchasing process of an available set, a consideration set, and selection is not reflected. Different criteria may be used for narrowing the available set to the consideration set and selecting from the consideration set. For example, the consideration set may be narrowed down based on the price and selected from the consideration set based on the design. Therefore, modeling selection from a consideration set is an important issue.

  The present invention has been made in view of the above circumstances, and a purchase data analysis apparatus and method capable of estimating a model for selecting a product to be purchased by a user based on the purchase behavior of the user when a consideration set is given. And to provide a program.

In order to achieve the above object, the purchase data analysis apparatus of the present invention provides an input purchase that is a set of purchase data including identification information for identifying a user and information indicating the type of product purchased by the user. Among the set of data sets, identification information for identifying each user, and available product types, an input including a consideration set that is a set of product types to be considered by the user is input. Product that each user purchases from the consideration set on the assumption that reading means for reading the consideration set data and the purchase data set and the consideration set data are not purchased based on the purchase data set and estimating means for estimating the model parameters of the model for selecting the output means and the including of outputting the model parameters estimated by the estimating means The purchase data analyzing apparatus, wherein the estimation means assigns each topic of the plurality of topics to each purchase data of the purchase data set, and calculates a topic assigned to each purchase data by the assignment means. A parameter for estimating a parameter representing the ease of selection of the product type from the consideration set in the topic for each combination of the topic and the product type so that the likelihood indicating likelihood is high. A set of topics assigned to each purchase data when the topic assignment by the assignment means and the parameter estimation by the parameter estimation means are repeated until a predetermined end condition is satisfied. And the parameters for each combination of the topic and the product type. A set of data, the estimation results of the model parameters.

  According to the purchase data analysis apparatus of the present invention, an input purchase data set which is a set of purchase data including identification information for identifying the user and information indicating the type of the product purchased by the user by the reading means. And an input consideration set including a consideration set that is a set of product types to be considered by the user among a set of identification information for identifying the user and available product types for each user. Read data and.

  A model for selecting products to be purchased by each user from the consideration set when it is assumed by the estimation means that the products not included in the consideration set are not purchased based on the purchase data set and the consideration set data. Estimate model parameters. The output unit outputs the model parameter estimated by the estimation unit.

  Thus, based on the purchase data set and the consideration set data, when it is assumed that the products not included in the consideration set are not purchased, the model parameters of the model for selecting the products to be purchased by each user from the consideration set are estimated. Thus, a model for selecting a product to be purchased by the user can be estimated based on the purchase behavior of the user when the consideration set is given.

Further, the purchase data analysis method of the present invention is an input purchase data set which is a set of purchase data including identification information for identifying a user and information indicating the type of product purchased by the user by the reading means. And an input consideration set including a consideration set that is a set of product types to be considered by the user among a set of identification information for identifying the user and available product types for each user. Each user purchases from the consideration set when it is assumed that a product not included in the consideration set is not purchased based on the purchase data set and the consideration set data by the step of reading data and the estimation means Estimating a model parameter of a model for selecting a product; and a model parameter estimated by the estimating means by an output means. A purchase data analyzing method comprising the steps of outputting over data, a step of estimating the model parameters by the estimating means, the assignment means and the purchase data of the purchase data set, any of a plurality of topics For each combination of the topic and the product type so that the likelihood indicating the likelihood of the topic assigned to each purchase data by the assigning means is increased by the step of assigning the topic and the parameter estimating means. Estimating a parameter indicating the ease of selection of the product type from the consideration set in the topic, and allocating the topic by the allocating means and estimating the parameter by the parameter estimating means Is repeated until a predetermined end condition is satisfied. Kino, a set of topics assigned to each purchase data, and said set of parameters for each combination of type of the topic and the product, the estimation results of the model parameters.

  Moreover, the purchase data analysis program of this invention is a program for functioning a computer as each means which comprises said purchase data analysis apparatus.

  As described above, according to the purchase data analysis apparatus, method, and program of the present invention, each product when it is assumed that a product not included in the consideration set is not purchased based on the purchase data set and the consideration set data. Estimating a model for selecting a product to be purchased by the user based on the purchase behavior of the user when the consideration set is given by estimating a model parameter of a model for the user to select a product to purchase from the consideration set The effect of being able to be obtained is obtained.

It is a block diagram which shows the functional structure of the purchase data analysis apparatus of the 1st Embodiment of this invention. It is a flowchart which shows the content of the purchase data analysis process routine in the purchase data analysis apparatus of the 1st Embodiment of this invention. It is a flowchart which shows the flow of the process which estimates the topic set and parameter set in the purchase data analysis apparatus of the 1st Embodiment of this invention. It is a figure which shows the accuracy comparison result of the purchase data analysis apparatus of this Embodiment, and another method. It is a figure which shows an example of topic extraction. It is a figure which shows an example of topic extraction.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
<Outline of model>
First, Table 1 shows the notation.

Among the available items set {i} i _{= 1} ^I, the considerations set is a product set of the consideration of the user u and _{s u = (s um) m} = 1 Mu. As a set to be considered, a set of products viewed in an online store, a set of products that meet user conditions, and the like can be considered. When a consideration set s _u is given, the probability that the user u selects the product i is expressed by the following equation (1).

Here, z∈ {1,..., K} is a topic, and P (z | u) represents the probability that the user u is interested in a product belonging to the topic z. P (i | z, s _u ) is a probability of selecting the product i when the topic z and the consideration set _su are given, and is represented by, for example, the following expression (2).

Here, λ _zi represents the ease of selection of the product i from the consideration set in topic z. The denominator on the right side of the above equation (2) is a normalization term necessary to satisfy the probability value property that the sum is 1.

As shown in the above equation (1), the purchased product is determined by the user's interest P (z | u) and the ease of selecting the product for each topic P (i | z, s _u ). The probability that a product not included in the consideration set is selected is zero.

  Here, the case of modeling based on a topic model has been explained as an example, but selection of products not included in the consideration set based on other purchased product selection models such as maximum entropy model and similarity-based collaborative filtering model The probability may be 0, and a normalization term may be added and modeled to obtain a probability value.

<System configuration>
A purchase data analysis apparatus 100 according to the first embodiment includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Read) that stores a program for executing a purchase data analysis processing routine to be described later. Only memory). This computer can be functionally represented by a configuration including a reading unit 10, a storage unit 20, an estimation unit 30, and an output unit 40, as shown in FIG.

  The reading unit 10 includes a consideration set reading unit 11 and a purchase set reading unit 12.

The consideration set reading unit 11 reads the input consideration set S = {s _u } _{u = 1} ^U of all users. The purchase set reading unit 12 reads purchase data sets X = {x _u } _{u = 1} ^U and x _u = (x _un ) _{n = 1} ^Nu for all users. Hereinafter, the nth purchase data _xun of the user u is also referred to as “purchase n of the user u”.

The storage unit 20 includes a consideration set storage unit 21 in which the read consideration set S = {s _u } ^U _{u = 1} of all the users is read, and a purchase data set X = {x _u } ^{U of} all the read users. A purchase set storage unit 22 storing _{u = 1,} a topic set storage unit 23 storing topic set Z, and a parameter set storage unit 24 storing parameter set Λ are provided.

  The estimation unit 30 includes an initialization unit 31, a topic estimation unit 32, and a parameter estimation unit 33.

The initialization unit 31 initializes the topic set Z by randomly assigning a topic to each purchase data of the purchase set read by the purchase set reading unit 12. The topic corresponds to the user's interest, for example, and indicates that the purchase n of the user u represented by the purchase data is interested in the product of the topic k. The topic set Z is Z = {z _u } _{u = 1} ^U (z _u = {z _un } _{n = 1} ^Nu ). The initialization unit 31 initializes each parameter of the parameter set by setting values at random. The parameter set is Λ = {λ _z } _{z = 1} ^Z (λ _z = {λ _zi } _{i = 1} ^I ).

  In addition, hyper parameters are also initialized appropriately. For example, the hyperparameter for the Dirichlet distribution is set to α = 1, and the hyperparameter for the normal distribution is set to β = 1.

The topic estimation unit 32 calculates the degree of attribution indicating to which topic each purchase n = 1,..., N _{u of} each user u = 1,. The degree of attribution needs to take the values of the user's topic ratio, merchandise purchase probability for each topic, and a consideration set as inputs. For example, the degree of belonging to which the purchase data _xun representing the purchase n of the user u belongs to the topic k is calculated by the following equation (3).

Here, \ un represents that the nth purchase of user u is excluded. N _uk represents the number of purchase data to which the topic k is assigned among the purchase data of the user u.

The topic estimation unit 32 assigns a new topic z _un based on the calculated degree of attribution. For example, the topic with the highest degree of attribution among topics belonging to each topic k (k = 1,..., K) of the purchase data _xun calculated by the above equation (3) is topic 1. If the is the topic 1, assigned as the topic of the purchase data x _un.

  Next, the parameter estimation unit 33 estimates the parameter set Λ so that the posterior probability of the parameter when data is given is maximized. The objective function to be maximized in the estimation of the parameter set Λ is expressed by the following equation (4).

  Here, the second term is a regularization term, and β represents the hyperparameter. By using an optimization method such as the quasi-Newton method, a global optimal solution that maximizes the objective function can be obtained.

  By repeating the topic estimation unit 32 and the parameter estimation unit 33 until the end condition is satisfied, the topic set Z and the parameter set Λ are estimated. The topic set Z and the parameter set Λ are examples of model parameters.

As termination conditions, for example, topic set Z = (z _u ) _{u = 1} ^U , z _u = (z _un ) _{n = 1} ^Nu , and parameter set Λ = (λ _z ) _{z = 1} ^Z , λ _z = ( The likelihood that the current estimate of λ _zi ) _{i = 1} ^I can explain the input data reasonably, that is, the likelihood converges, which represents the likelihood of the topic assigned to each purchase data. As characteristics of likelihood, (a) the likelihood increases when the number of topic types allocated to purchase data of a certain user is small, and (b) the product of purchase data to which a certain topic is assigned is a specific product. When it is a set, the likelihood needs to be high. Note that a value having an equivalent property such as a posteriori probability can be used instead of the likelihood.

  Further, as the termination condition, it can be used that the number of repetitions, the calculation time, etc. reach a predetermined value.

The output unit 40 includes a topic ratio output unit 41 and a parameter output unit 42. The topic ratio output unit 41 calculates the topic ratio θ _uk for each combination of the user u and the topic k according to the following equation (5) using the estimated topic set Z, and outputs the calculated topic ratio set. .

Further, the topic ratio output unit 41 uses the calculated topic ratio set to perform clustering of users with similar purchasing behavior and outputs a clustering result. For example, by assigning each user u to the topic k having the highest topic ratio θ _uk , user clustering is performed, and the user clustering result (user group assigned to each topic k) is output.

The parameter output unit 42 outputs the estimated parameter set Λ. The parameter output unit 42 uses the parameter set Λ to cluster products that are easily purchased at the same time, and outputs a clustering result. For example, by assigning each product i to the topic k having the highest parameter λ _ki , the products are clustered, and the product clustering result (the product group assigned to each topic k) is output.

Further, the parameter output unit 42 uses the topic ratio set, the parameter set Λ, and the consideration set S to obtain a product (recommended product) suitable for the user's purchase behavior for each user, and recommends each user. Output the product. For example, by substituting the set of topic ratios θ _uk , parameter set Λ, and consideration set S into the above equation (1) (where P (z | u) = θ _uz ), The probability of purchasing is estimated, and for each user, a product with a high probability is obtained as a recommended product.

<Operation of purchasing data analyzer>
When the consideration set S of each user and the purchase data set X of all users are input to the purchase data analysis apparatus 100 of this embodiment, the purchase data analysis processing routine shown in FIG. Executed.

  First, in step S <b> 101, the input consideration set S and purchase data set X are read by the reading unit 10 and stored in the consideration set storage unit 21 and the purchase set storage unit 22.

  In the next step S102, the estimation unit 30 estimates a topic set Z and a parameter set Λ for each purchase data in the purchase data set.

  Here, step S102 is realized by the processing routine shown in FIG.

  In step S110, the initialization unit 31 initializes the topic set Z by randomly assigning the topic z to each purchase data x of the purchase data set X, and randomly sets each parameter of the parameter set Λ. Is assigned a value to initialize the parameter set Λ. The initialized topic set Z is stored in the topic set storage unit 23, and the initialized parameter set Λ is stored in the parameter set storage unit 24. Also, appropriate values are set for the hyper parameters α and β.

  In the next step S111, a variable u for identifying the user is set to 1 which is an initial value, and in step S112, a variable n for identifying purchase is set to 1 which is an initial value.

In step S113, the topic estimation unit 32 uses the topic set Z and parameter set Λ initialized in step S110 or the previously estimated topic set Z and parameter set Λ for the purchase data _xun . According to the above equation (3), the degree of attribution of the purchase data _xun to each topic k is calculated, a new topic is assigned to the purchase data _xun based on the calculated degree of attribution, and the contents of the topic set storage unit 23 Update.

In the next step S114, it is determined whether or not the variable n is greater than or equal to N _u . If the variable n is less than N _u , n is incremented by 1 in step S115, and the above-mentioned number of steps S113 is performed. Return to. Meanwhile, the variable n is, when it is more than N _u determines that assigned the topic for all purchase data of the user u, the process proceeds to step S116.

  In step S116, it is determined whether or not the variable u is greater than or equal to a constant U indicating the number of users. If the variable u is less than U, u is incremented by 1 in step S117, and the above-mentioned number step. Return to S112. On the other hand, if the variable u is greater than or equal to U, it is determined that a topic has been assigned to purchase data for all users, and the process proceeds to step S118.

  In step S118, a global optimum solution that maximizes the objective function shown in the above equation (4) is obtained using the topic set Z obtained in step S113, and the parameter set Λ obtained as the solution is obtained as the parameter set Λ. The contents of the parameter set storage unit 24 are updated.

  In step S119, it is determined whether or not a repetitive end condition is satisfied. For example, it is determined whether or not the likelihood calculated based on the topic set Z obtained in step S113 has converged. If the likelihood has not converged, the process returns to step S111 and the processes of steps S111 to S118 are repeated. On the other hand, if it is determined that the likelihood has converged, the process of step S102 is terminated. The topic set Z finally stored in the topic set storage unit 23 is the estimation result of the topic set, and the parameter set Λ finally stored in the parameter set storage unit 24 is the estimation result of the parameter set. is there.

  Next, in step S103 of FIG. 2, the topic ratio output unit 41 calculates and outputs the topic ratio set according to the above equation (5) using the topic set Z estimated in step S102. Also, the topic ratio output unit 41 clusters users using the calculated topic ratio set, and outputs the user clustering result.

  In step S104, the parameter output unit 42 outputs the parameter set Λ estimated in step S102. Further, the parameter output unit 42 clusters the products using the parameter set Λ and outputs the clustering result of the products. The parameter output unit 42 calculates the purchase probability of each product for each user using the parameter set Λ, determines the recommended product, outputs the recommended product for each user, and ends the process.

  As described above, according to the purchase data analysis apparatus according to the first embodiment, based on the purchase data set and the consideration set of each user, products not included in the consideration set are assumed to be not purchased. Estimating a model for selecting a product to be purchased by a user based on the purchase behavior of the user when a consideration set is given by estimating model parameters of a model for each user to select a product to purchase from the consideration set be able to.

  In addition, it is possible to estimate a more precise model than the conventional technique that does not use a consideration set, and it is possible to improve prediction accuracy.

  Further, by analyzing the purchase data, it is possible to recommend clustering of products that are easily purchased at the same time, clustering of users having similar purchasing behavior, and products suitable for the purchasing behavior of the user.

[Second Embodiment]
Next, a second embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The second embodiment is different from the first embodiment in that the probability that each product is included in the consideration set is input as the consideration set data.

The consideration set reading unit 11 of the purchase data analyzing apparatus according to the second embodiment receives a set of probabilities q _u = {q for each user u included in the consideration set, instead of the consideration set S 1. _ui } _{i = 1} ^I is read and stored in the consideration set storage unit 21.

Here, when a set of probabilities q _u = {q _ui } _{i = 1} ^I in which each product of the user u is included in the consideration set, the probability that the user u selects the product i is expressed by the above equation (1). Instead of, it is expressed by the following equation (6).

Here, q _ui represents the probability that the product i is included in the consideration set of the user u. The probability set q may be estimated from data.

The parameter output unit 42 substitutes a topic ratio set, a parameter set Λ, and a set q _u of probabilities that each product u is included in the consideration set in the above equation (6) (where P (z | u) = θ _uz ), the probability of purchasing each product of each user is estimated, and for each user, a product with a high probability is obtained as a recommended product.

  In addition, about the other structure and effect | action of the purchase data analyzer which concerns on 2nd Embodiment, since it is the same as that of 1st Embodiment, description is abbreviate | omitted.

<Example>
In order to evaluate the present invention, an experiment using purchase data in an online store was conducted. The effectiveness of the proposed method was confirmed using three purchase history data from November 19, 2010 to February 28, 2011, March 31, and May 9. Table 2 below shows the number of users, the number of brands, the total number of purchases, and the total number of considerations (page view) of each purchase history data. Users who did not purchase more than once were omitted. In addition, the product viewed by the user in the online store was considered as a consideration set. The brand is the type of product represented by the purchase data.

  For 10% of users, one randomly selected purchase was used as test data, the other purchases were used as learning data, and the purchase brand included in the test data was predicted using the learning data. The correct answer rate at that time was used as an evaluation index. The results are shown in FIGS. As a comparison method, a normal topic model (LDA: Latent Dirichlet Allocation) that does not use a consideration set was used. In LDA, the product with the highest purchase probability in the set to be considered is set as the predicted product. The present invention (proposed) has a high percentage of correct answers in all data sets, and it has been found that a purchase brand (purchased product) can be predicted more accurately by incorporating a consideration set.

  Moreover, the topic extraction result at the time of experimenting by dividing goods by fashion and cosmetics is shown in FIGS. As a topic extraction result, for each topic, a set of products having a high appearance probability is shown in a set of products of purchase data to which the topic is assigned. FIG. 5 shows the results of experiments performed on products classified into fashion, and FIG. 6 shows the results of experiments performed on products classified into cosmetics. From the above, it was found that the same topic was assigned to related products.

  The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  In the above embodiment, the method of estimating the topic set and the parameter set by alternately repeating the Gibbs sampling of the topic and the full likelihood maximization regarding the parameter has been described. However, the present invention is not limited to this. For example, it is possible to estimate the topic set and parameter set using Markov chain Monte Carlo method, variational Bayes method, posterior probability maximization method, etc.

  Moreover, although the above-mentioned purchase data analysis apparatus has a computer system inside, if a "computer system" is using the WWW system, it shall include a homepage provision environment (or display environment). .

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Reading part 20 Storage part 30 Estimation part 31 Initialization part 32 Topic estimation part 33 Parameter estimation part 40 Output part 41 Topic ratio output part 42 Parameter output part 100 Purchasing data analysis apparatus

Claims (7)

Identification information for identifying the user, input purchase data set that is a set of purchase data including information indicating the type of product purchased by the user, and identification information for identifying the user for each user, Reading means for reading input consideration set data including a consideration set which is a set of product types to be considered by the user among a set of available product types;
Based on the purchase data set and the consideration set data, a model parameter of a model for selecting a product to be purchased by each user from the consideration set when it is assumed that a product not included in the consideration set is not purchased is estimated. An estimation means;
Output means for outputting the model parameters estimated by the estimation means;
A purchasing data analysis device including a,
The estimation means includes
Assigning means for assigning any one of a plurality of topics to each purchase data of the purchase data set;
The product from the consideration set in the topic for each combination of the topic and the product type so that the likelihood indicating the likelihood of the topic assigned to each purchase data by the assigning means is high. Parameter estimation means for estimating a parameter indicating the ease of selection of the type, and
A set of topics assigned to each purchase data, the topic, and the product when the topic assignment by the assignment means and the parameter estimation by the parameter estimation means are repeated until a predetermined end condition is satisfied. A purchase data analyzing apparatus that uses the set of parameters for each combination of types as an estimation result of the model parameters.   The output means performs at least one of user clustering for a user group with similar purchasing behavior and product clustering for a product group that is easily purchased using the model parameter estimated by the estimating unit, The purchase data analysis apparatus according to claim 1, wherein the clustering result is output. The assigning means uses the set of topics assigned to each purchase data, the set of parameters for each combination of the topic and the product type, and the consideration set of each user, and calculating a degree of membership indicating a probability of belonging to the topic, based on the membership, purchase data analyzing apparatus according to claim 1 or 2, wherein allocating the topic to the purchasing data. In order to identify the user for each user and the input purchase data set which is a set of purchase data including identification information for identifying the user and information indicating the type of product purchased by the user by the reading means Reading the input consideration set data including a consideration set that is a set of product types to be considered by the user among the set of identification information and available product types;
A model of a model in which each user selects a product to be purchased from the consideration set based on the purchase data set and the consideration set data, assuming that the product not included in the consideration set is not purchased. Estimating the parameters;
Outputting a model parameter estimated by the estimating means by an output means;
A purchasing data analysis methods, including,
The step of estimating the model parameter by the estimating means includes
Assigning any one of a plurality of topics to each purchase data of the purchase data set by an assigning means;
In the set of considerations in the topic for each combination of the topic and the product type, the parameter estimation unit increases the likelihood indicating the likelihood of the topic allocated to each purchase data by the allocation unit. Estimating a parameter representing the ease of selection of the product type from
A set of topics assigned to each purchase data, the topic, and the product when the topic assignment by the assignment means and the parameter estimation by the parameter estimation means are repeated until a predetermined end condition is satisfied. A purchase data analysis method using the set of parameters for each combination of types as an estimation result of the model parameters. The step of outputting by the output means includes at least clustering of users for obtaining a user group having similar purchasing behavior using the model parameter estimated by the estimating means and clustering of products for obtaining a product group that is easily purchased at the same time. The purchase data analysis method according to claim 4 , wherein one is performed and the clustering result is output.   The step of assigning by the assigning means uses the set of topics assigned to each purchase data, the set of parameters for each combination of the topic and the product type, and the consideration set of each user. The purchase data analysis method according to claim 4 or 5, wherein a degree of attribution indicating a probability that the data belongs to each topic is calculated, and the topic is assigned to the purchase data based on the degree of attribution. The purchase data analysis program for functioning a computer as each means which comprises the purchase data analysis apparatus of any one of Claims 1-3 .

Images